Nine famous websites, zero agent-ready

Last week, we launched a free tool at /tools/agent-ready. Enter a URL, and it will score the site’s public surface for the agent era: there are fifteen checks across four categories, and no signup or installation is required. Before writing this, we tested it on nine of the most famous websites on the internet — companies with world-class engineering teams. The average score was 37 out of 100, and not one of them was deemed agent-ready.

Here are the scores from September 29: GitHub scored 45, making it the best of the group. The New York Times followed with 44, Harvard scored 40, OpenAI had 39, Shopify scored 36, Stripe came in at 35, IBM and Mattel both scored 30, while Anthropic (the company behind Claude) also scored 30.

01What the check actually does

The tool fetches a site as a plain, logged-out HTTP agent would, which means it does not use cookies, JavaScript rendering, or any authenticated access. This is increasingly how your site gets read. Agents cannot make assumptions — if the machine-readable layer is missing, they improvise, and that improvisation becomes what your customers receive as information.

The tool consists of fifteen checks across four key groups, all fairly weighted:

• —Discovery basics: This includes essential elements like robots.txt, a declared sitemap, and accurate metadata — basic requirements that every site should meet.
• —Content accessible to agents: This involves markdown negotiation and an llms.txt / agents.md index, ensuring that what you offer is formatted in a way that agents can quote rather than just approximate.
• —Bot policy: This includes explicit rules for AI crawlers and content signals, so that the intent is clearly stated rather than assumed.
• —The agentic surface: This focuses on emerging action protocols, including MCP, A2A, skills, API catalogs, authentication, and bot signing.

A site earns the full weight of a check for passing, half for a warning, and nothing for a failure.

02Interpreting the scores

An obvious objection arises: how can Anthropic, which builds agents, score only 30? There are two honest explanations.

First, the fourth category is still emerging and carries significant weight. It's normal for sites to fail under the current protocols, as very few are publishing that layer yet; the tool reflects this reality rather than glossing over it. Most of the gap between a 37 and a perfect score resides in this category.

Second, that's the purpose of a measurement tool. The term "mobile-friendly" was once considered specialized language; now it's standard. By publishing this tool, we highlight gaps that need addressing. And to clarify, a perfect score indicates that a site is discoverable and machine-negotiable — not that agents will appreciate the content. That part remains your responsibility.

03We tested our own site

Since we offer this service, we tested our own site, thethirdwaycompany.com, before launching the tool. It successfully passed the exam, showcasing features like llms.txt, agent guidance, a machine-readable bot policy, and the emerging protocols. The tool also suggests fixes in a format that a coding agent can implement — rather than presenting a PDF that ends up filed away. This check is free regardless.

If you prefer assistance with this, we provide that service, typically within a week, and will score the site with the same tool before and after.

04The five-minute version

Run your domain through the checker. Review the failures, provide the suggested fixes to your coding agent, and reassess in the next quarter. For a companion test, select your most essential product and have two AI assistants compare it against a competitor. Any discrepancies they find reveal how agents are currently framing your offering. We previously discussed this in "Your Next Customer Might Be an AI Agent."

Your next customer might be a person, but their initial visit increasingly occurs through an agent. The central question the checker poses is whether that visitor can access and understand your content.